Abstract

Multi-word sequences (MWSs) are processed more efficiently than non-formulaic language and support L2 oral production (Conklin & Schmitt, 2012), yet fluency research has overwhelmingly focused on monologic speech (Hougham et al., 2024). This remains a critical gap given that lexical bundles (LBs) are central to academic interaction, where speakers coordinate turns and negotiate meaning in real time (Mauranen, 2012). Furthermore, while MWSs associate with oral fluency across proficiency levels (Tavakoli & Uchihara, 2020), the relationship between bundle length and fluency in multilogues (interactions involving three or more speakers) remains unexplored. This study examines how LB length and frequency predict L2 fluency in academic multilogues across proficiency levels among L1 Japanese learners of English.

Speech data were elicited from 120 L1 Japanese (L2 English) participants grouped by CEFR level (A2–C1; n = 30 per level) based on an a priori power analysis. Participants completed interactive academic multilogue tasks. Audio and transcripts were analyzed using a structured five-phase framework combining acoustic measurement in PRAAT with a retrospective, nine-category interactional timeline (distinguishing categories such as own pauses, listener pauses, transition gaps, overlaps, and shared silence). To ensure data integrity during turn-taking analysis, simultaneous speech transitions were explicitly treated as distinct overlap events rather than conflated with zero-latency turn latencies. Following Hougham et al. (2024), LBs were evaluated using text-internal and text-external approaches: TAALES 2.0 identified shorter LBs against a spoken reference corpus (COCA), while AntConc identified four-word bundles. Longer bundles were manually refined against clause boundaries. Linear mixed-effects models assessed LB–fluency relationships across proficiency tiers.

Drawing on evidence that longer MWSs reduce cognitive planning load (Kormos, 2006), findings address how bundle length differentially impacts breakdown fluency and speech rate within dynamic conversational floors. By integrating corpus tools with interactional phonetics, this paper advances understanding of formulaic language in spoken multilogues.